U.S. patent application number 13/992614 was filed with the patent office on 2013-09-26 for attribute value estimation device, attribute value estimation method, program, and recording medium.
This patent application is currently assigned to TOKYO INSTITUTE OF TECHNOLOGY. The applicant listed for this patent is Yasuyuki Ihara, Masashi Sugiyama, Kazuya Ueki. Invention is credited to Yasuyuki Ihara, Masashi Sugiyama, Kazuya Ueki.
Application Number | 20130254143 13/992614 |
Document ID | / |
Family ID | 46206974 |
Filed Date | 2013-09-26 |
United States Patent
Application |
20130254143 |
Kind Code |
A1 |
Ueki; Kazuya ; et
al. |
September 26, 2013 |
ATTRIBUTE VALUE ESTIMATION DEVICE, ATTRIBUTE VALUE ESTIMATION
METHOD, PROGRAM, AND RECORDING MEDIUM
Abstract
The present invention provides an attribute value estimation
device capable of yielding highly accurate estimation results even
when people from multiple races are estimation targets. The
attribute value estimation device for estimating, from data input
thereto, an attribute value of the data includes: a data
acquisition unit (1) that acquires data for which an attribute
value is to be estimated; a discrete quantity estimation unit (2)
that estimates the attribute value as a discrete quantity based on
the data acquired by the data acquisition unit (1) and in
accordance with a previously learned determination criterion; a
first LSPC (3) that estimates the attribute value as a discrete
quantity based on data input from the discrete quantity estimation
unit (2); and an integration unit (4) that integrates a first
discrete quantity estimation value estimated by the discrete
quantity estimation unit (2) and a second discrete quantity
estimation value estimated by the first LSPC (3).
Inventors: |
Ueki; Kazuya; (Koto-ku,
JP) ; Ihara; Yasuyuki; (Koto-ku, JP) ;
Sugiyama; Masashi; (Meguro-ku, JP) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
Ueki; Kazuya
Ihara; Yasuyuki
Sugiyama; Masashi |
Koto-ku
Koto-ku
Meguro-ku |
|
JP
JP
JP |
|
|
Assignee: |
TOKYO INSTITUTE OF
TECHNOLOGY
Meguro-ku, Tokyo
JP
NEC SOFT, LTD.
Koto-ku, Tokyo
JP
|
Family ID: |
46206974 |
Appl. No.: |
13/992614 |
Filed: |
November 17, 2011 |
PCT Filed: |
November 17, 2011 |
PCT NO: |
PCT/JP2011/076516 |
371 Date: |
June 7, 2013 |
Current U.S.
Class: |
706/12 ; 706/15;
706/46 |
Current CPC
Class: |
G06N 7/005 20130101;
G06N 5/02 20130101; G06N 3/0427 20130101; G06Q 30/02 20130101; G06K
9/6278 20130101; G06K 2009/00322 20130101; G06K 9/6292 20130101;
G06K 9/00288 20130101 |
Class at
Publication: |
706/12 ; 706/46;
706/15 |
International
Class: |
G06N 5/02 20060101
G06N005/02 |
Foreign Application Data
Date |
Code |
Application Number |
Dec 8, 2010 |
JP |
2010-273829 |
Claims
1. An attribute value estimation device for estimating, from data
input thereto, an attribute value of the data, the attribute value
estimation device comprising: a data acquisition unit that acquires
data for which an attribute value is to be estimated; at least one
estimation unit selected from: a discrete quantity estimation unit
that estimates the attribute value as a discrete quantity; and a
continuous quantity estimation unit that estimates the attribute
value as a continuous quantity, the estimation unit estimating the
attribute value based on the data acquired by the data acquisition
unit and in accordance with a previously learned determination
criterion; and a least-squares probabilistic classifier that
estimates the attribute value as a discrete quantity based on data
input from the estimation unit; and an integration unit that
integrates an estimation value estimated by the estimation unit and
a discrete quantity estimation value estimated by the least-squares
probabilistic classifier.
2. The attribute value estimation device according to claim 1,
further comprising: a scoring unit that scores the estimation value
estimated by the estimation unit; and a discrete quantity scoring
unit that scores the discrete quantity estimation value estimated
by the least-squares probabilistic classifier, wherein the
integration unit integrates a first score value obtained by the
scoring unit and a second score value obtained by the discrete
quantity scoring unit.
3. The attribute value estimation device according to claim 1,
wherein the integration unit integrates the estimation value, the
discrete quantity estimation value, and the score values with a
weight being assigned to at least one of the estimation value, the
discrete quantity estimation value, and the score values.
4. The attribute value estimation device according to claim 1,
wherein the least-squares probabilistic classifier previously
learns the determination criterion, and in the learning of the
determination criterion, the least-squares probabilistic classifier
calculates a kernel function only when a class of an input feature
quantity is the same as a correct class to which a training sample
belongs.
5. The attribute value estimation device according to claim 1,
wherein the least-squares probabilistic classifier previously
learns the determination criterion, and in the learning of the
determination criterion, the center of kernel is placed in a class
for which the number of training samples is the smallest.
6. The attribute value estimation device according to claim 1,
wherein at least one estimation unit selected from the discrete
quantity estimation unit and the continuous quantity estimation
unit comprises a neural network, dimensionality reduction of the
data acquired by the data acquisition unit is performed by the
neural network, and the attribute value is estimated based on the
dimensionality-reduced data, and the least-squares probabilistic
classifier estimates the attribute value as a discrete quantity
based on the dimensionality-reduced data.
7. The attribute value estimation device according to claim 1,
wherein the data acquired by the data acquisition unit is face
image data, and the attribute value is a face attribute value.
8. The attribute value estimation device according to claim 7,
wherein the face attribute value is at least one attribute value
selected from the group consisting of age group, age, gender, and
race.
9. An attribute value estimation method for estimating, from input
data, an attribute value of the data, the attribute value
estimation method comprising: a data acquisition step of acquiring
data for which an attribute value is to be estimated; an estimation
step of estimating the attribute value as at least one of a
discrete quantity and a continuous quantity based on the data
acquired in the data acquisition step and in accordance with a
previously learned determination criterion; a discrete quantity
estimation step of estimating the attribute value as a discrete
quantity based on data processed in the estimation step; and an
integration step of integrating an estimation value estimated in
the estimation step and a discrete quantity estimation value
estimated in the discrete quantity estimation step, wherein a
least-squares probabilistic classifier is used in the discrete
quantity estimation step.
10. The attribute value estimation method according to claim 9,
further comprising: a scoring step of scoring the estimation value
estimated in the estimation step; and a discrete quantity scoring
step of scoring the discrete quantity estimation value estimated by
the least-squares probabilistic classifier in the discrete quantity
estimation step, wherein, in the integration step, a first score
value obtained in the scoring step and a second score value
obtained in the discrete quantity scoring step are integrated.
11. The attribute value estimation method according to claim 9,
wherein, in the integration step, the estimation value, the
discrete quantity estimation value, and the score values are
integrated with a weight being assigned to at least one of the
estimation value, the discrete quantity estimation value, and the
score values.
12. The attribute value estimation method according to claim 9,
wherein the least-squares probabilistic classifier previously
learns the determination criterion, and in the learning of the
determination criterion, the least-squares probabilistic classifier
calculates a kernel function only when a class of an input feature
quantity is the same as a correct class to which a training sample
belongs.
13. The attribute value estimation method according to claim 9,
wherein the least-squares probabilistic classifier previously
learns the determination criterion, and in the learning of the
determination criterion, the center of kernel is placed in a class
for which the number of training samples is the smallest.
14. The attribute value estimation method according to claim 9,
wherein in the estimation step, at least one of the discrete
quantity and the continuous quantity is estimated using a neural
network, dimensionality reduction of the data acquired in the data
acquisition step is performed by the neural network, and the
attribute value is estimated based on the dimensionality-reduced
data, and in the discrete quantity estimation step, the
least-squares probabilistic classifier estimates the attribute
value as a discrete quantity based on the dimensionality-reduced
data.
15. The attribute value estimation method according to claim 9,
wherein the data acquired in the data acquisition step is face
image data, and the attribute value is a face attribute value.
16. The attribute value estimation method according to claim 15,
wherein the face attribute value is at least one attribute value
selected from the group consisting of age group, age, gender, and
race.
17. A program that causes a computer to execute the attribute value
estimation method according to claim 9.
18. A recording medium having recorded thereon the program
according to claim 17.
Description
TECHNICAL FIELD
[0001] The present invention relates to an attribute value
estimation device, an attribute value estimation method, a program,
and a recording medium.
BACKGROUND ART
[0002] Conventional attribute value estimation devices for
estimating, from data input thereto, an attribute value of the data
include those that estimate an attribute value of input data by
extracting features of the data and then comparing the
thus-extracted features with features of training samples that the
devices have learned previously. In such devices, estimation
results may be treated as discrete quantities (Patent Document 1)
or as continuous quantities (Patent Document 2).
CITATION LIST
Patent Document(s)
[0003] Patent Document 1: JP 2007-58828 A
[0004] Patent Document 2: JP 2005-148880 A
SUMMARY OF THE INVENTION
Problem to be Solved by the Invention
[0005] In estimation of a face attribute value such as an age by
the above-described devices, the devices can yield highly accurate
estimation results when people from a specific race, such as
Japanese, are estimation targets. However, when estimation targets
are face images of people from multiple races including various
facial features, training samples might be biased, so that it is
difficult to achieve the same level of accuracy as that achieved
when the estimation targets are people from a specific race.
[0006] With the foregoing in mind, it is an object of the present
invention to provide an attribute value estimation device, an
attribute value estimation method, a program, and a recording
medium, with which highly accurate estimation results can be
obtained even when people from multiple races are estimation
targets.
Means for Solving Problem
[0007] In order to achieve the above object, the present invention
provides an attribute value estimation device for estimating, from
data input thereto, an attribute value of the data, including:
[0008] a data acquisition unit that acquires data for which an
attribute value is to be estimated;
[0009] at least one estimation unit selected from: a discrete
quantity estimation unit that estimates the attribute value as a
discrete quantity; and a continuous quantity estimation unit that
estimates the attribute value as a continuous quantity, the
estimation unit estimating the attribute value based on the data
acquired by the data acquisition unit and in accordance with a
previously learned determination criterion; and
[0010] a LSPC (Least-Squares Probabilistic Classifier) that
estimates the attribute value as a discrete quantity based on data
input from the estimation unit; and
[0011] an integration unit that integrates an estimation value
estimated by the estimation unit and a discrete quantity estimation
value estimated by the least-squares probabilistic classifier.
[0012] The present invention also provides an attribute value
estimation method for estimating, from input data, an attribute
value of the data, including:
[0013] a data acquisition step of acquiring data for which an
attribute value is to be estimated;
[0014] an estimation step of estimating the attribute value as at
least one of a discrete quantity and a continuous quantity based on
the data acquired in the data acquisition step and in accordance
with a previously learned determination criterion;
[0015] a discrete quantity estimation step of estimating the
attribute value as a discrete quantity based on data processed in
the estimation step; and
[0016] an integration step of integrating an estimation value
estimated in the estimation step and a discrete quantity estimation
value estimated in the discrete quantity estimation step,
[0017] wherein a LSPC is used in the discrete quantity estimation
step.
[0018] The present invention also provides a program that causes a
computer to execute the attribute value estimation method according
to the present invention.
[0019] The present invention also provides a recording medium
having recorded thereon the program according to the present
invention.
Effects of the Invention
[0020] According to the present invention, it is possible to obtain
highly accurate estimation results even when people from multiple
races are estimation targets.
BRIEF DESCRIPTION OF DRAWINGS
[0021] FIG. 1 is a block diagram showing an example (Embodiment 1)
of the attribute value estimation device of the present
invention.
[0022] FIG. 2 is a block diagram showing another example
(Embodiment 2) of the attribute value estimation device of the
present invention.
[0023] FIG. 3 is a block diagram showing still another example
(Embodiment 3) of the attribute value estimation device of the
present invention.
[0024] FIG. 4 is a block diagram showing still another example
(Embodiment 4) of the attribute value estimation device of the
present invention.
[0025] FIG. 5 is a block diagram showing still another example
(Embodiment 5) of the attribute value estimation device of the
present invention.
[0026] FIG. 6 is a block diagram showing still another example
(Embodiment 7) of the attribute value estimation device of the
present invention.
[0027] FIGS. 7A and 7B are each a graph showing the relationship
between age and the standard deviation of estimation error.
[0028] FIG. 8 shows graphs showing the distributions of score
vector components before and after integration.
MODE FOR CARRYING OUT THE INVENTION
[0029] The attribute value estimation device of the present
invention preferably is configured so that it further includes: a
scoring unit that scores the estimation value estimated by the
estimation unit; and a discrete quantity scoring unit that scores
the discrete quantity estimation value estimated by the LSPC,
wherein the integration unit integrates a first score value
obtained by the scoring unit and a second score value obtained by
the discrete quantity scoring unit. Similarly, the attribute value
estimation method of the present invention preferably is configured
so that it further includes: a scoring step of scoring the
estimation value estimated in the estimation step; and a discrete
quantity scoring step of scoring the discrete quantity estimation
value estimated by the LSPC in the discrete quantity estimation
step, wherein, in the integration step, a first score value
obtained in the scoring step and a second score value obtained in
the discrete quantity scoring step are integrated.
[0030] The attribute value estimation device of the present
invention preferably is configured so that the integration unit
integrates the estimation value, the discrete quantity estimation
value, and the score values with a weight being assigned to at
least one of the estimation value, the discrete quantity estimation
value, and the score values. Similarly, the attribute value
estimation method of the present invention preferably is configured
so that, in the integration step, the estimation value, the
discrete quantity estimation value, and the score values are
integrated with a weight being assigned to at least one of the
estimation value, the discrete quantity estimation value, and the
score values.
[0031] The attribute value estimation device and attribute value
estimation method according to the present invention preferably are
configured so that the LSPC previously learns the determination
criterion, and in the learning of the determination criterion, the
LSPC calculates a kernel function only when a class of an input
feature quantity is the same as a correct class to which a training
sample belongs. With this configuration, it is possible to further
speed up calculations at the time of learning, for example.
[0032] The attribute value estimation device and attribute value
estimation method according to the present invention preferably are
configured so that the LSPC previously learns the determination
criterion, and in the learning of the determination criterion, the
center of kernel is placed in a class for which the number of
training samples is the smallest. With this configuration, it is
possible to further speed up calculations at the time of learning,
for example.
[0033] The attribute value estimation device of the present
invention preferably is configured so that at least one estimation
unit selected from the discrete quantity estimation unit and the
continuous quantity estimation unit includes a neural network,
dimensionality reduction of the data acquired by the data
acquisition unit is performed by the neural network, and the
attribute value is estimated based on the dimensionality-reduced
data, and the least-squares probabilistic classifier estimates the
attribute value as a discrete quantity based on the
dimensionality-reduced data. Similarly, the attribute value
estimation method of the present invention preferably is configured
so that, in the estimation step, at least one of the discrete
quantity and the continuous quantity is estimated using a neural
network, dimensionality reduction of the data acquired in the data
acquisition step is performed by the neural network, and the
attribute value is estimated based on the dimensionality-reduced
data, and in the discrete quantity estimation step, the LSPC
estimates the attribute value as a discrete quantity based on the
dimensionality-reduced data.
[0034] The attribute value estimation device and attribute value
estimation method according to the present invention preferably are
configured so that the data acquired by the data acquisition unit
and the data acquired in the data acquisition step are face image
data, and the attribute value is a face attribute value.
[0035] The attribute value estimation device and attribute value
estimation method according to the present invention preferably
configured so that the face attribute value is at least one
attribute value selected from the group consisting of age group,
age, gender, and race.
[0036] Next, the attribute value estimation device, attribute value
estimation method, program, and recording medium according to the
present invention will be described with reference to illustrative
examples. It is to be noted, however, that the present invention is
by no means limited to the following examples. In FIGS. 1 to 6 to
be described below, the same components are given the same
reference numerals.
Embodiment 1
[0037] FIG. 1 shows a block diagram of an attribute value
estimation device of the present embodiment. The attribute value
estimation device of the present embodiment has a discrete quantity
estimation unit as the estimation unit, and can be used for
estimating an attribute value such as race or gender, for example.
As shown in FIG. 1, the attribute value estimation device of the
present embodiment includes, as main components, a data acquisition
unit 1, a discrete quantity estimation unit 2, a first LSPC 3, and
an integration unit 4. Examples of the data acquisition unit 1
include image acquisition units such as CCD (Charge Coupled Device)
cameras, CMOS (Complementary Metal Oxide Semiconductor) cameras,
and image scanners. The discrete quantity estimation unit 2 stores
previously learned determination criteria. Each of the discrete
quantity estimation unit 2, the first LSPC 3, and the integration
unit 4 can be any dedicated hardware (e.g., a central processing
unit (CPU) or the like), or can be realized on a computer by
software processing, for example.
[0038] The discrete quantity estimation unit 2 extracts a feature
quantity used for attribute value estimation from input data. Using
the feature quantity extracted from the input data and the
determination criteria, the discrete quantity estimation unit 2
estimates an attribute value of the input data as a discrete
quantity. In the case where the attribute value is a race, for
example, the discrete quantity may be white, black, yellow,
Mongoloid, or mixed (biracial or multiracial), for example. In the
case where the attribute value is a gender, the discrete quantity
may be male or female, for example.
[0039] The discrete quantity estimation unit 2 can extract the
feature quantity from the input data using a conventionally known
method, examples of which include edge extraction and
binarization.
[0040] The discrete quantity estimation unit 2 can estimate the
attribute value as a discrete quantity from the feature quantity in
accordance with the determination criteria using a conventionally
known method, examples of which include: the use of a neural
network, a Gaussian mixture model, or a support vector machine;
linear discrimination analysis; logistic regression analysis; and a
k-nearest neighbor classification method.
[0041] The first LSPC 3 estimates the attribute value as a discrete
quantity based on, as a new feature quantity, data input from the
discrete quantity estimation unit 2. The first LSPC 3 solves a
posterior probability model in each class using a squared loss.
Thus, the most distinctive feature of the first LSPC 3 is that it
can achieve ultra-high speed learning. Besides, the first LSPC 3
models the posterior probability in the form of density ratio, so
that it also has a feature that it is resistant to imbalance in the
number of pieces of data among respective classes of training
samples. For example, when people from multiple races are
estimation targets, it is difficult to collect training samples
evenly for various classifications such as age group, race, and
gender. Thus, this feature of the first LSPC 3 is advantageous when
the estimation targets are people from multiple races.
[0042] The first LSPC 3 estimates the posterior probability
distribution p (y|x) of an attribute class y regarding an input
feature quantity (facial feature quantity) x in the form of density
ratio represented by the following Expression (1). Examples of the
attribute class include age group classes, gender classes, and race
classes.
p ( y | x ) = p ( x , y ) p ( x ) ( 1 ) ##EQU00001## [0043] p(x):
probability distribution of training samples [0044] p(x,y): joint
probability distribution of training samples In the end, the
attribute class with the highest posterior probability (the left
side of the following Expression (2)) is set to an estimated
attribute class.
[0044] y ^ = arg min y p ( y | x ) ( 2 ) ##EQU00002##
[0045] The first LSPC 3 learns the posterior probability p (y|x)
using a squared loss. This allows, for example, the learning time
to be reduced to one several hundredth while maintaining the
pattern recognition accuracy equivalent to those achieved by
conventional methods.
[0046] Moreover, since the posterior probability is estimated in
the form of density ratio represented by Expression (1), the
estimation result is less susceptible to the influence of imbalance
in the number of pieces of training sample data among respective
classes (e.g., the number of pieces of training sample data in a
particular class is small).
[0047] Next, the least square fitting of the posterior probability
will be described. First, the posterior probability of the
attribute class y is modeled using the following linear model.
q ( y | x ; .alpha. ) = i = 1 b .alpha. I .phi. I ( x , y ) ( 3 )
##EQU00003## [0048] {.alpha..sub.l}.sub.l=1.sup.b: parameter [0049]
{.phi..sub.l (x,y)}.sub.l=1.sup.b: basis function that is
non-negative
[0050] The first LSPC 3 learns a parameter cc (the following
expression) in such a manner that the following square error
J.sub.0 is minimized.
.alpha. = ( .alpha. 1 , , .alpha. I ) T J 0 ( .alpha. ) = 1 2
.intg. y = 1 c ( q ( y | x ; .alpha. ) - p ( y | x ) ) 2 p ( x ) x
= 1 2 .intg. y = 1 c ( q ( y | x ; .alpha. ) ) 2 p ( x ) x - .intg.
y = 1 c q ( y | x ; .alpha. ) p ( x , y ) x + 1 2 .intg. y = 1 c (
p ( y | x ) ) 2 p ( x ) x ( 4 ) ##EQU00004##
The last term in the above expression is a constant and thus can be
ignored. By approximating the expectation value J in the first two
terms in the above expression by the sample mean, the following
Expression (5) is obtained.
J ^ ( .alpha. ) = 1 2 .alpha. T H .alpha. - h ^ T .alpha. ( 5 ) { H
^ = 1 n i = 1 n y = 1 c .phi. ( x i , y ) .phi. ( x i , y ) T h ^ =
1 n i = 1 n .phi. ( x i , y i ) ( 5 a ) ##EQU00005##
An l.sub.2-regularization term is added in order to prevent
overfitting, thus yielding the following unconstrained optimization
problem.
.alpha. ^ = arg min .alpha. [ 1 2 .alpha. T H ^ .alpha. - h ^ T
.alpha. + .lamda. 2 .alpha. T .alpha. ] ( 6 ) ##EQU00006##
The solution of Expression (6) is analytically given by the
following Expression (7).
{tilde over (.alpha.)}=(H+.lamda. I.sub.b).sup.-1h (7) [0051]
I.sub.b: b-dimensional unit matrix Parameter correction is
performed in accordance with the following Expression (8) so as to
make all the parameters non-negative, thus ensuring the
non-negativity of the posterior probability.
[0051] {circumflex over (.alpha.)}.sub.l=max(0,.alpha..sub.l) for
l=1,2, . . . , b (8)
Finally, normalized correlation is performed so as to make the sum
of all the classes equal to 1, thus obtaining the solution of the
posterior probability.
p ^ ( y | x ) = .alpha. T .phi. ( x , y ) y ' = 1 c .alpha. T .phi.
( x , y ' ) ( 9 ) ##EQU00007##
Alternatively, instead of the processes represented by Expressions
(8) and (9), the following process may be performed to obtain the
solution of the posterior probability.
p _ ( y | x ) = { 1 Z max ( 0 , .alpha. T .phi. ( x , y ) if Z >
0 1 c otherwise Z = y ' = 1 c max ( 0 , .alpha. T .phi. ( x , y ' )
) ##EQU00008##
[0052] When estimating the discrete quantity of the attribute
value, the discrete quantity estimation unit 2 and the first LSPC 3
may each output the discrete quantity estimation value accompanying
stochastic representation of the estimation result. The discrete
quantity estimation value accompanying stochastic representation of
the estimation result may be as follows, for example: in the case
where the attribute value is a race, the discrete quantity
estimation value may be, for example, "the estimation target is
white with a probability of 80% and black with a probability of
20%"; and in the case where the attribute value is a gender, the
discrete quantity estimation value may be, for example, "the
estimation target is male with a probability of 80% and female with
a probability of 20%". With this configuration, in the case where
it is estimated that the estimation target is female from part of
its appearance but there is a possibility that the estimation
target may be male according to any other determination criteria
(e.g., the estimation target is a person with long hair), it is
possible to output an estimation result with higher accuracy.
[0053] The integration unit 4 integrates a first discrete quantity
estimation value estimated by the discrete quantity estimation unit
2 and a second discrete quantity estimation value estimated by the
first LSPC 3. The integration unit 4 outputs, as an estimation
result, an attribute class obtained after the integration. In the
present embodiment, since the first discrete quantity estimation
value and the second discrete quantity estimation value are
integrated, there is no risk that the estimation accuracy of any
particular attribute class might be low.
[0054] Furthermore, in the present embodiment, in order to further
improve the learning speed of the first LSPC 3, at least one of (1)
introduction of a delta kernel and (2) placement of the center of a
kernel may be performed, for example.
[0055] First, (1) introduction of a delta kernel will be described.
When the following training samples are given, the posterior
probability model (Expression (3)) of each attribute class y is
designed as represented by the following expression.
{(x.sub.i,y.sub.i)}.sub.i=1.sup.l [0056] x.sub.i: objective
variable (facial feature quantity) [0057] y.sub.i: explanatory
variable (attribute class)
[0057] q ( y | x ; .alpha. ) = y ' = 1 c I = 1 n .alpha. i ( y ' )
( x , x i , y , y i , y ' ) ##EQU00009## [0058] c: the number of
attribute classes [0059] n: the number of learning samples [0060]
K: kernel function determined by input feature quantity x, class y,
similarity of training samples {(x.sub.i,y.sub.i)}.sub.i=1.sup.l,
and class y' [0061] the number of parameters
.alpha..sub.i.sup.(y'): cn [0062] calculation amount required to
obtain analytic solution (Expression (7) above):
O(c.sup.3n.sup.3)
[0063] At the time of learning, a "delta kernel" for calculating
the kernel function is introduced only when a class of the input
feature quantity x is the same as a correct class to which a
training sample (in the present embodiment, the objective variable
(facial feature quantity)) x.sub.i belongs.
K'(x,x.sub.i,y,y.sub.i,y')=K(x,x.sub.i).delta..sub.y,y' (10)
.delta..sub.y,y' is a Kronecker delta represented by the following
Expression (11).
.delta. y , y ' = { 1 if y = y ' 0 otherwise ( 11 )
##EQU00010##
This reduces the number of parameters (cn), thus turning the matrix
in Expression (5a) to a block diagonal matrix for each attribute
class. In this case, the calculation amount required to obtain the
analytic solution (Expression (7)) is O(cn.sup.3).
[0064] Next, (2) placement of the center of a kernel will be
described. In the class y, the value of the posterior probability p
(y|x) is high in a region where the number of training samples is
large whereas it is almost 0 (zero) in a region where the number of
training samples is small. Therefore, the center of the kernel may
be placed where there are training samples in the class.
[0065] This makes the block of the matrix of Expression (5a) still
smaller, thus allowing the calculation amount for the inverse
matrix to be reduced further.
Embodiment 2
[0066] FIG. 2 shows another configuration of the attribute value
estimation device of the present invention. As shown in FIG. 2,
this attribute value estimation device has the same configuration
as the attribute value estimation device shown in FIG. 1, except
that it further includes a first discrete quantity scoring unit 5
and a second discrete quantity scoring unit 6.
[0067] The attribute value estimation device and an attribute value
estimation method according to the present embodiment will be
described more specifically with reference to an example where the
attribute value is a race, and the race estimation is performed for
three classes, namely, white, Asian, and black.
[0068] The first discrete quantity scoring unit 5 scores a first
discrete quantity estimation value estimated by the discrete
quantity estimation unit 2, and outputs, as a first score value,
certainty factors in the stochastic form (white: p.sub.1, Asian:
p.sub.2, black: p.sub.3).
[0069] The second discrete quantity scoring unit 6 scores a second
discrete quantity estimation value estimated by the first LSPC 3,
and outputs, as a second score value, certainty factors in the same
form as described above (white: q.sub.1, Asian: q.sub.2, black:
q.sub.3).
[0070] The integration unit 4 assigns weights .omega..sub.1 and
.omega..sub.2 determined in Embodiment 6 to be described below to
the first score value and the second score value, respectively, and
adds the thus-weighted first and second score values. Then, the
integration unit 4 outputs, as an estimation result, the race
having a highest score value among the following score values
R.sub.1, R.sub.2, and R.sub.3.
Score value for white:
R.sub.1=.omega..sub.1p.sub.1+.omega..sub.2q.sub.1
Score value for Asian:
R.sub.2=.omega..sub.1p.sub.2+.omega..sub.2q.sub.2
Score value for black:
R.sub.3=.omega..sub.1p.sub.3+.omega..sub.2q.sub.3
Embodiment 3
[0071] FIG. 3 shows still another example of the attribute value
estimation device of the present invention. The attribute value
estimation device of the present embodiment has a continuous
quantity estimation unit as the estimation unit, and can be used
for estimating an attribute value such as age group or age, for
example. As shown in FIG. 3, this attribute value estimation device
has the same configuration as the attribute value estimation device
shown in FIG. 1, except that it includes a continuous quantity
estimation unit 7 and a second LSPC 8 instead of the discrete
quantity estimation unit 2 and the first LSPC 3. The continuous
quantity estimation unit 7 stores previously learned determination
criteria.
[0072] The continuous quantity estimation unit 7 extracts a feature
quantity used for attribute value estimation from input data. Using
the feature quantity extracted from the input data and the
determination criteria, the continuous quantity estimation unit 7
estimates an attribute value of the input data as a continuous
quantity. The continuous quantity estimation unit 7 can extract the
feature quantity from the input data using a conventionally known
method, examples of which include edge extraction and
binarization.
[0073] The continuous quantity estimation unit 7 can estimate the
attribute value from the feature quantity in accordance with the
determination criteria using a conventionally known method,
examples of which include: the use of a neural network; multiple
regression analysis; support vector regression; kernel regularized
weighted least squares; and a k-nearest neighbor classification
method.
[0074] The second LSPC 8 estimates the attribute value as a
discrete quantity based on, as a new feature quantity, data input
from the continuous quantity estimation unit 7, in the same manner
as the above-described first LSPC 3.
[0075] The integration unit 4 integrates a continuous quantity
estimation value estimated by the continuous quantity estimation
unit 7 and a third discrete quantity estimation value estimated by
the second LSPC 8. The integration unit 4 outputs, as an estimation
result, an attribute value (e.g., an age group, an age, or the
like) obtained after the integration.
[0076] The attribute value estimation device of the present
embodiment may be configured so that, for example: the continuous
quantity estimation unit 7 includes a neural network; the second
LSPC 8 outputs certainty factors for respective age group classes
based on, as a new facial feature quantity, data whose
dimensionality has been reduced by the neural network (e.g.,
intermediate 100-dimensional data); and an age group with the
highest certainty factor is set to an estimated age group. In the
case where the discrete quantity estimation unit 2 in Embodiments 1
and 2 includes a neural network as in the present example, the
neural network may reduce the dimensionality of data as in the
present example.
Embodiment 4
[0077] FIG. 4 shows still another example of the attribute value
estimation device of the present invention. The attribute value
estimation device of present embodiment has both a discrete
quantity estimation unit and a continuous quantity estimation unit
as the estimation units, and can be used for estimating an
attribute value such as age group or age, for example. As shown in
FIG. 4, this attribute value estimation device has the same
configuration as the attribute value estimation device shown in
FIG. 1, except that it further includes a continuous quantity
estimation unit 7 and a second LSPC 8. The discrete quantity
estimation unit 2 and the first LSPC 3 are as described in
Embodiment 1. The continuous quantity estimation unit 7 and the
second LSPC 8 are as described in Embodiment 3.
[0078] The integration unit 4 integrates a first discrete quantity
estimation value estimated by the discrete quantity estimation unit
2, a second discrete quantity estimation value estimated by the
first LSPC 3, a continuous quantity estimation value estimated by
the continuous quantity estimation unit 7, and a third discrete
quantity estimation value estimated by the second LSPC 8. The
integration unit 4 outputs, as an estimation result, an attribute
value (e.g., an age group, an age, or the like) obtained after the
integration.
Embodiment 5
[0079] FIG. 5 shows still another example of the attribute value
estimation device of the present invention. As shown in FIG. 5,
this attribute value estimation device has the same configuration
as the attribute value estimation device shown in FIG. 4, except
that it further includes a first discrete quantity scoring unit 5,
a second discrete quantity scoring unit 6, a continuous quantity
scoring unit 9, and a third discrete quantity scoring unit 10. The
first discrete quantity scoring unit 5 and the second discrete
quantity scoring unit 6 are as described in Embodiment 2.
[0080] Scoring in the present embodiment will be described with
reference to an example where the continuous quantity estimation
unit 7 includes a neural network. In this example, the continuous
quantity estimation unit 7 and the second LSPC 8 assign scores to
the respective ages from 1 to 70 (at 1-year intervals), and scores
are thus output in the vector form.
[0081] First, (1) scoring of an output from the neural network will
be described. The neural network used in the present example has
been trained through regression model learning so that it can
perform age estimation at 1-year intervals. Thus, an output
therefrom is in the form of a single scalar y (age). An output from
the neural network is scored in the following manner, with
consideration given to the fact that human age perception
characteristics are nonuniform (Kazuya UEKI, Masashi SUGIYAMA, and
Yasuyuki IHARA, "Omomitsuki-kaiki ni yoru ningen no chikaku-tokusei
wo kouryo sita nenrei-suitei (Age Estimation Considering Human
Perception Characteristic by Weighted Regression)", Proceedings of
the 15th Symposium on Sensing via Image Information (SSII09), no.
IS4-23 (CD-ROM), Yokohama, 2009, 6. 10-12).
[0082] In this example, the mean value of perceived ages (apparent
ages) of the same subject estimated by a plurality of estimators is
set to a "real age". The standard deviation of estimation error in
age perception with respect to the real age y is represented as
.omega..sub.age(y). The .omega..sub.age(y) is nonuniform, as shown
in FIGS. 7A and 7B. More specifically, while error in age
perception for younger subjects is small, error in age perception
for older subjects is great. FIG. 7A shows the standard deviation
of error when the ages of female subjects at the respective ages
were estimated from their face images, and FIG. 7B shows the
standard deviation of error when the ages of male subjects at the
respective ages were estimated from their face images.
[0083] Using this .omega..sub.age(.cndot.) (".cndot." is the
underlined part in the following expression), an output age from
the neural network (the underlined part in the following
expression) is scored in the following manner.
output score : f 1 = { f 1 ( z ) } z = 1 70 respective components
of f 1 : f 1 ( z ) = 1 2 .pi. .omega. age ( y ~ _ ) exp ( - ( z - y
_ ~ ) 2 - 2 .omega. age ( y ~ _ ) 2 ) ( z = 1 , , 70 ) ( 12 )
##EQU00011##
[0084] At this time, as shown in the upper left graph of FIG. 8,
the distribution of the components of the score f.sub.1 is in the
form of normal distribution exhibiting small dispersion for younger
subjects and large dispersion for older subjects, with the peak
being formed when the output age is equal to the underlined part in
the above expression. That is, scoring is performed so as to
reflect the fact that there is a low probability that an estimated
age for a younger subject might be an age around the real age of
the subject while there is a high probability that an estimated age
for an older subject might be an age around the real age of the
subject.
[0085] Next, (2) scoring of an output from the second LSPC 8 will
be described. The second LSPC 8 outputs certainty factors for the
respective age groups in the stochastic form. In the present
example, the scoring is performed so that an output from the second
LSPC 8 is in the same form as the score f.sub.1 assigned to the
above-described output from the neural network.
[0086] For example, when age group estimation is performed for
three classes, namely, an early age group (0 to 19 years old), a
middle age group (20 to 49 years old), and an old age group (over
50 years old), it is assumed that the certainty factors output from
the second LSPC 8 are as follows: the early age group: p.sub.1, the
middle age group: p.sub.2, and the old age group: p.sub.3.
[0087] At this time, a score is temporarily assigned to each age z
(z=1, . . . 70) in accordance with the following Expression
(13).
{circumflex over (f)}.sub.2(z)=p.sub.i (13)
Thereafter, the score is normalized using the following
expression.
f 2 ( z ) = f ^ 2 ( z ) z = 1 70 f ^ 2 ( z ) ( 14 )
##EQU00012##
Then, the following expression is set to an output score from the
second LSPC 8.
f.sub.2={f.sub.2(z)}.sub.z=1.sup.70
[0088] The upper right graph of FIG. 8 shows an image of the
distribution of components of the output score vector from the
second LSPC 8.
[0089] In the present embodiment, the integration unit 4 integrates
a first score value obtained by the first discrete quantity scoring
unit 5, a second score value obtained by the second discrete
quantity scoring unit 6, a third score value obtained by the
continuous quantity scoring unit 9, and a fourth score value
obtained by the third discrete quantity scoring unit 10. The
integration unit 4 outputs, as an estimation result, an age
obtained after the integration.
Embodiment 6
[0090] The present embodiment is carried out using the attribute
value estimation device shown in FIG. 5. The present embodiment is
the same as Embodiment 5, except that the integration unit 4
performs the above-described integration with weighs being assigned
to the third score value and the fourth score value.
[0091] In the present embodiment, weighting of output scores is
carried out in the following manner, for example. Weights
.omega..sub.1 and .omega..sub.2 are assigned respectively to scores
obtained by the continuous quantity scoring unit 9 and the third
discrete quantity scoring unit 10 regarding each age z
(1.ltoreq.z.ltoreq.70, at 1-year intervals), and the thus-weighted
scores are added to each other. The method for determining optimal
weights will be described below.
F = .omega. 1 f 1 + .omega. 2 f 2 { F = { F ( z ) } z = 1 70 F ( z
) = .omega. 1 f 1 ( z ) + .omega. 2 f 2 ( z ) ( 15 )
##EQU00013##
[0092] Then, an age group to which an age z* that satisfies
z*=argmax.sub.z{F(z)} belongs is set to an output age group from
the integration unit 4. The lower graph in FIG. 8 shows an image of
the distribution of score vector components after the
integration.
[0093] Next, the method for searching for the weights will be
described. Using validation data (data that is not used in model
learning), optimal weights .omega..sub.1 and .omega..sub.2 are
searched for one by one thoroughly. More specifically, the
evaluation of the integration unit 4 is performed using the
validation data under the following conditions: numerical widths of
.omega..sub.1 and .omega..sub.2: 0 to 1, search intervals: 0.01. A
score having the highest score (the mean value of recognition rates
in each category) when evaluated using the validation data is
employed as an optimal weight.
Embodiment 7
[0094] FIG. 6 shows still another example of the attribute value
estimation device of the present invention. As shown in FIG. 6,
this attribute value estimation device includes a data acquisition
unit 1, a race estimation unit 11, an age group estimation unit 21
for whites, an age group estimation unit 22 for Asians, an age
group estimation unit 23 for blacks, and an integration unit 4. The
race estimation unit 11 includes the discrete quantity estimation
unit 2, first LSPC 3, first discrete quantity scoring unit 5, and
second discrete quantity scoring unit 6 of the attribute value
estimation device shown in FIG. 2. The age group estimation unit 21
for whites, the age group estimation unit 22 for Asians, and the
age group estimation unit for blacks 23 each include the discrete
quantity estimation unit 2, first LSPC 3, first discrete quantity
scoring unit 5, second discrete quantity scoring unit 6, continuous
quantity estimation unit 7, second LSPC 8, continuous quantity
scoring unit 9, and third discrete quantity scoring unit 10 of the
attribute value estimation device shown in FIG. 5.
[0095] The attribute value estimation device and an attribute value
estimation method according to the present embodiment will be
described more specifically with reference to an example where the
race estimation unit 11 performs race estimation for three classes,
namely, white, Asian, and black.
[0096] The race estimation unit 11 outputs certainty factors in the
stochastic form (white: p.sub.1, Asian: p.sub.2, black: p.sub.3) as
a score value.
[0097] The age group estimation unit 21 for whites, the age group
estimation unit 22 for Asians, and the age group estimation unit 23
for blacks output, as score values at each age z
(1.ltoreq.z.ltoreq.70, at 1-year intervals), the following
Expressions (16) to (18), respectively.
score value for white: W={W(z)}.sub.z=1.sup.70 (16)
score value for Asian: A={A(z)}.sub.z=1.sup.70 (17)
score value for black: B={B(z)}.sub.z=1.sup.70 (18)
[0098] Using the certainty factors (in the stochastic form) output
from the race estimation unit 11, the integration unit 4 integrates
the score values for each age z (1.ltoreq.z.ltoreq.70, at 1-year
intervals) as shown below. The integration unit 4 outputs, as an
age group estimation result including the race estimation, an age
group to which an age z* that satisfies z*=argmax.sub.z{G(z)}
belongs.
G = p 1 W + p 2 A + p 3 B ##EQU00014## { G = { G ( z ) } z = 1 70 G
( z ) = p 1 W ( z ) + p 2 A ( z ) + p 3 B ( z ) ##EQU00014.2##
[0099] While the present invention has been described above with
reference to illustrative embodiments, the present invention is by
no means limited thereto. Various changes and modifications that
may become apparent to those skilled in the art may be made in the
configuration and specifics of the present invention without
departing from the scope of the present invention.
[0100] This application claims priority from Japanese Patent
Application No. 2010-273829 filed on Dec. 8, 2010. The entire
disclosure of this Japanese patent application is incorporated
herein by reference.
EXPLANATION OF REFERENCE NUMERALS
[0101] 1 data acquisition unit [0102] 2 discrete quantity
estimation unit [0103] 3 first LSPC (Least-Squares Probabilistic
Classifier) [0104] 4 integration unit [0105] 5 first discrete
quantity scoring unit [0106] 6 second discrete quantity scoring
unit [0107] 7 continuous quantity estimation unit [0108] 8 second
LSPC (Least-Squares Probabilistic Classifier) [0109] 9 continuous
quantity scoring unit [0110] 10 third discrete quantity scoring
unit [0111] 11 race estimation unit [0112] 21 age group estimation
unit for whites [0113] 22 age group estimation unit for Asians
[0114] 23 age group estimation unit for blacks
* * * * *